-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
144 lines (126 loc) · 5.61 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
import os
import tqdm
import math
import logging
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as transforms
from torch.autograd import Variable
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import StepLR
from torch.optim.lr_scheduler import MultiStepLR
from torch.utils.tensorboard import SummaryWriter
from dataloader.data_utils import *
from dataloader.samplers import *
from methods.cosine_classifier import CosClassifier
from utils.utils import *
from utils.fsl_inc import *
from sync_batchnorm import convert_model
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# experiments arguments
parser.add_argument('--dataroot', type=str, default='PATH/CEC-CVPR2021/data/')
parser.add_argument('--dataset', type=str, default='mini-imagenet')
parser.add_argument('--method', type=str, default='imprint')
parser.add_argument('--base_mode', type=str, default='avg_cos')
parser.add_argument('--norm_first', action='store_true')
parser.add_argument('--exp_dir', type=str, default='experiment')
# training arguments
parser.add_argument('--epoch', type=int, default=120)
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--batch_size_new', type=int, default=0)
parser.add_argument('--batch_size_test', type=int, default=100)
parser.add_argument('--init_lr', type=float, default=-1)
parser.add_argument('--schedule', type=str, default='Milestone', choices=['Step', 'Milestone'])
parser.add_argument('--milestones', nargs='+', type=int, default=-1)
parser.add_argument('--step', type=int, default=40)
parser.add_argument('--gamma', type=float, default=0.1)
parser.add_argument('--val_start', type=int, default=50)
parser.add_argument('--val_interval', type=int, default=5)
parser.add_argument('--change_val_interval', type=int, default=70)
parser.add_argument('--num_workers', type=int, default=8)
parser.add_argument('--report_binary', action='store_true')
args = parser.parse_args()
args = set_up_datasets(args)
args.norm_first = True
if args.init_lr == -1:
if args.dataset == 'cifar100':
args.init_lr = 0.1
elif args.dataset == 'mini_imagenet':
args.init_lr = 0.1
elif args.dataset == 'cub200':
args.init_lr = 0.01
else:
Exception('Undefined dataset name!')
if args.milestones == -1:
if args.dataset == 'cifar100':
args.milestones = [120, 160]
elif args.dataset == 'mini_imagenet':
args.milestones = [120, 160]
elif args.dataset == 'cub200':
args.milestones = [50, 70, 90]
else:
Exception('Undefined dataset name!')
args.checkpoint_dir = '%s/%s' %(args.exp_dir, args.dataset)
if not os.path.isdir(args.checkpoint_dir):
os.makedirs(args.checkpoint_dir)
logging.basicConfig(filename=os.path.join(args.checkpoint_dir, 'train.log'), level=logging.INFO)
logging.info(args)
print(args)
# init model
model = CosClassifier(args, phase='pre_train')
model.cuda()
# init optimizer
optimizer = torch.optim.SGD(model.parameters(), lr=args.init_lr, momentum=0.9, weight_decay=5e-4, nesterov=True)
if args.schedule == 'Step':
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.step, gamma=args.gamma)
elif args.schedule == 'Milestone':
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.milestones, gamma=args.gamma)
loss_fn = nn.CrossEntropyLoss()
# dataset in pre-training phase
trainset, trainloader, testloader = model.get_dataloader(0)
print(len(trainset))
# training
best_test_acc_base = 0
best_test_epoch_base = 0
for epoch in range(args.epoch):
if epoch >= args.change_val_interval:
args.val_interval = 1
# torch.cuda.empty_cache()
model.train()
if args.schedule != 'Step' and args.schedule != 'Milestone':
adjust_learning_rate(optimizer, epoch, init_lr=args.init_lr, n_epoch=args.epoch)
tqdm_gen = tqdm.tqdm(trainloader)
loss_avg = 0
for i, X in enumerate(tqdm_gen):
data, label = X
data = data.cuda()
label = label.cuda()
pred = model(flag='base_forward', input=data)
loss = loss_fn(pred, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
tqdm_gen.set_description('e:%d loss = %.4f' % (epoch, loss.item()))
loss_avg += loss.item()
if args.schedule == 'Step' or args.schedule == 'Milestone':
lr_scheduler.step()
out_str = '======epoch: %d avg loss: %.6f======'%(epoch, loss_avg/len(trainloader))
print(out_str)
logging.info(out_str)
# testing
model.eval()
if (epoch == 0) or (epoch > args.val_start and (epoch+1) % args.val_interval == 0):
acc_list = model.test_inc_loop(epoch=epoch)
test_acc_base = acc_list[0]
if test_acc_base > best_test_acc_base:
best_test_acc_base = test_acc_base
best_test_epoch_base = epoch
outfile = os.path.join(args.checkpoint_dir, 'best_model_%s.tar'%(args.base_mode))
torch.save({'epoch':epoch, 'state':model.state_dict()}, outfile)
out_str = '==========Epoch: %d Best Base Test acc = %.2f%%==========='%(best_test_epoch_base, 100*best_test_acc_base)
print(out_str)
logging.info(out_str)